Roman Garnett

never seen a robot

[Roman Garnett @glottalstop]

@maosbot Related: why don't we always use the optimal (i.e., full lookahead) policy? Because it's so expensive to compute. (Not to mention that the model is probably wrong, as mentioned by another response, so the extra cost is probably not even worth it.)

[Roman Garnett @glottalstop]

@maosbot My "certainly not!" was in reference only to "...[for all problems]." For cheap experiments we should expect cheap policies to on the frontier, especially as "smarter" policies often come with costs that may be orders of magnitude greater.

[Roman Garnett @glottalstop]

@maosbot And yes, algorithmic development has an inherent cost as well that I should account for when making decisions. It could be perfectly rational to adopt a suboptimal randomized procedure for one problem and concentrate your efforts where they may make a bigger impact.

[Roman Garnett @glottalstop]

@maosbot I was not commenting on the efficiency of RNGs per se, rather that computation has an inherent cost, and that using randomized procedures may actually represent a rational decision when it maximizes expected cost-adjusted progress on your problem.

[Roman Garnett @glottalstop]

@maosbot the five papers in that session have received a total of 1002 citations with a median count of 39; perhaps there's something to the idea. food for thought...

[Roman Garnett @glottalstop]

@maosbot oh, and the session chairs were also the ICML program chairs...I'm guessing they had a reason to be late ;)

[Roman Garnett @glottalstop]

@maosbot from Léon Bottou: https://t.co/yMLm9WaAcq "The impact of the papers published in this session remains to be seen. However we are proud to report that the session was packed. Since attendees and reviewers are the same persons, one wonder why these papers received such criticism."

[Roman Garnett @glottalstop]

@maosbot from the same ancient email thread: [Mike] Whoa! Awes. Not bad for three days work or however long it took us. I have my Erdos number!!! [Roman] i am actually utterly stunned by the acceptance btw

[Roman Garnett @glottalstop]

@cse361 a 785

[Roman Garnett @glottalstop]

@mdreid it's the sort of optimization problem where elucidating the utility function itself is hard.

[Roman Garnett @glottalstop]

@mdreid ah thanks, was missing your replies to others

[Roman Garnett @glottalstop]

@mdreid recommended papers?

[Roman Garnett @glottalstop]

@bkkbrad should have been an anchor in that url taking you to talk #384

[Roman Garnett @glottalstop]

@clydeiii sigh ijltg

[Roman Garnett @glottalstop]

just got EXTREMELY burned by icml session chairs

[Roman Garnett @glottalstop]

icml session: "food for thought"...here comes our paper

[Roman Garnett @glottalstop]

awesome dataset: predict color of beef given storage conditions